Top story
GLM-5.3 shows how much post-training can move a fixed base
https://huggingface.co/zai-org/GLM-5.3
GLM-5.3 keeps GLM-5.2’s base model and attributes every improvement to post-training. Z.ai reports Terminal Bench 3.0 rising from 4.6 to 28.3, DeepSWE from 46.2 to 66.9, and an 84.5 CyberGym score. The model is open-weight, with vLLM, SGLang, and Transformers support. I would not give it an instant SOTA crown: these are vendor-reported numbers, and several evaluations use huge contexts plus six-to-ten-hour budgets. What matters to me is how far a fixed base moved on long-horizon coding and cyber tasks. HN commenter scosman says it “feels like Opus 4.8”; I take that as a reason to evaluate, not validation. If you can host it, run GLM-5.3 this week on real repositories with your existing agent harness and a fixed spend. Cost, latency, and completion rate should decide whether these open weights beat your current API—not the boldest cell in Z.ai’s benchmark table.
Two changes worth checking
The OpenAI Python SDK moved to HTTPX2. Default API calls still work, but the SDK no longer installs httpx or certifi, and TLS now uses the operating-system trust store. I would check minimal containers, corporate TLS proxies, custom transports, and RESPX tests before upgrading. simonw frames the fork as a stability path while HTTPX approaches 1.0. Read the migration guide before bumping.
AI enforcement knocked Luanti off Google Play. Tracer.AI filed a Microsoft DMCA notice without identifying a specific asset; a similar 2023 notice took 46 days to unwind. Hnrobert42 identifies the broken incentive: false positives are nearly free while false negatives cost the enforcer. I read Luanti’s account as an operations warning about automation without accountable human review.
One day, post-training expands model capability while an AI agent removes an app without evidence: capability is outrunning accountability.
— Tin